My hypothesis is that since PicBreeder can learn abstract concepts like symmetry, maybe (just maybe), a similar neural network could learn an abstract concept like language (yes, I know, language is a lot more complex than symmetry). Both PicBreeder and FishNet use something called a CPPN (Compositional Pattern Producing Network), which uses a different architecture than what we know as an LLM. You can find the full paper for PicBreeder at https://wiki.santafe.edu/images/1/1e/Secretan_ecj11.pdf (no, I haven’t read the whole thing either).
If you’re interested in helping me out, just go to FishNet and click the sequence you find the most interesting, and if you find something cool, like a word, a recognizable structure, or anything else, click the “I think I found something cool” button! If you were wondering: it's called FishNet because in early testing I had it learn to output “fish fish fish fish fish fish it”.
Source code’s here: https://github.com/Z-Coder672/FishNet/tree/main/code
*Not sure about the trustworthiness of this unofficial PicBreeder site, I wouldn’t click that save button, but here’s the link anyway: https://nbenko1.github.io/. The official site at picbreeder.org is down :(